About
The C4 database and classifier
This website accompanies our paper describing the generation of the C4 database and its utilization to build a cerebellar cell-type classifier: Beau et al., 2025.
The C4 database is a dataset consisting of ground-truth recordings of different types of cerebellar neurons obtained using Neuropixels probes in awake mice. The ground-truth identification was established using the optotagging approach, which combines optogenetic activation of specific cell types (using the appropriate Cre lines) with pharmacological synaptic blockade to confirm direct optogenetically-triggered responses. Layer information was determined using Phyllum, a custom analysis tool that accurately identifies cerebellar layers from electrophysiological recordings. The database includes carefully curated recordings of 202 neurons: 69 Purkinje cells simple spike units, 58 Purkinje cells complex spike units, 27 molecular layer interneurons, 18 Golgi cells, and 30 mossy fibers.
The classifier uses a semi-supervised deep learning approach that combines unsupervised pre-training with supervised classification. It first uses variational autoencoders trained on a large unlabeled dataset (3,090 neurons) to reduce the dimensionality of both spike waveforms and three-dimensional autocorrelograms (3D-ACGs). These compressed representations, along with layer information, serve as inputs to a supervised classifier trained on the ground-truth dataset. The classifier generates confidence-calibrated predictions by averaging outputs across model initializations. As describer in the paper, when requiring a confidence ratio above 2 (ratio of highest to second-highest predicted probability), the classifier achieves >95% accuracy for predicting cell type on the ground-truth dataset. It also generalizes well to expert-classified recordings from both mice and monkeys, demonstrating its ability to be used across species and experimental setups.
The cerebellum cell-types classification collaboration (C4)
The Cerebellum Cell-types Classification Collaboration (C4) is a multi-institutional effort to create a database of cerebellar cell-types and use it to build a cell-type classifier for general use in the cerebellar neuroscience community. The name also refers to the involvement of teams across four universities (Häusser and Clark lab at University College London, Cohen at Bar-Ilan University, the Lisberger and Hull labs at Duke University, and the Medina lab at Baylor College of Medicine).
For any questions or feedback, please contact us.